Revert "fix: add cache flush as dependency for bcs ccs synchronization"

This reverts commit 5e57bb2a32.

Signed-off-by: Compute-Runtime-Validation <compute-runtime-validation@intel.com>

opencl/source/command_queue/enqueue_common.h

@@ -580,7 +580,6 @@ BlitProperties CommandQueueHw<GfxFamily>::processDispatchForBlitEnqueue(CommandS
         blitProperties.csrDependencies.timestampPacketContainer.push_back(&timestampPacketDependencies.cacheFlushNodes);
         blitProperties.csrDependencies.timestampPacketContainer.push_back(&timestampPacketDependencies.previousEnqueueNodes);
         blitProperties.csrDependencies.timestampPacketContainer.push_back(&timestampPacketDependencies.barrierNodes);
-        blitProperties.csrDependencies.timestampPacketContainer.push_back(&timestampPacketDependencies.multiCsrDependencies);
     }
     blitProperties.multiRootDeviceEventSync = multiRootDeviceEventSync;
     auto currentTimestampPacketNode = timestampPacketContainer->peekNodes().at(0);
@@ -1438,14 +1437,7 @@ cl_int CommandQueueHw<GfxFamily>::enqueueBlit(const MultiDispatchInfo &multiDisp
     if (isCacheFlushForBcsRequired() && gpgpuSubmission) {
         timestampPacketDependencies.cacheFlushNodes.add(allocator->getTag());
     }
-    for (auto &dependentCsr : csrDeps.csrWithMultiEngineDependencies) {
+
-        auto tag = allocator->getTag();
-        timestampPacketDependencies.multiCsrDependencies.add(tag);
-        bool submitStatus = dependentCsr->submitDependencyUpdate(tag);
-        if (!submitStatus) {
-            return CL_OUT_OF_RESOURCES;
-        }
-    }
     obtainNewTimestampPacketNodes(1, timestampPacketDependencies.previousEnqueueNodes, clearAllDependencies, bcsCsr);
     csrDeps.timestampPacketContainer.push_back(&timestampPacketDependencies.previousEnqueueNodes);
 

opencl/source/helpers/properties_helper.cpp

@@ -58,14 +58,13 @@ void EventsRequest::fillCsrDependenciesForTimestampPacketContainer(CsrDependenci
         if (pushDependency) {
             csrDeps.timestampPacketContainer.push_back(timestampPacketContainer);
 
-            const auto &productHelper = event->getCommandQueue()->getDevice().getProductHelper();
             if (!sameCsr) {
+                const auto &productHelper = event->getCommandQueue()->getDevice().getProductHelper();
                 if (productHelper.isDcFlushAllowed()) {
                     if (!dependentCsr->isLatestTaskCountFlushed()) {
                         flushDependentCsr(*dependentCsr, csrDeps);
                         currentCsr.makeResident(*dependentCsr->getTagAllocation());
                     }
-                    csrDeps.csrWithMultiEngineDependencies.insert(dependentCsr);
                 }
             }
         }

opencl/test/unit_test/command_queue/blit_enqueue_1_tests.cpp

@@ -1373,11 +1373,11 @@ HWTEST_TEMPLATED_F(BlitEnqueueTaskCountTests, givenBlockedEventWhenWaitingForCom
     userEvent.setStatus(CL_COMPLETE);
 
     clWaitForEvents(1, &outEvent2);
-    EXPECT_EQ(3u, ultGpgpuCsr->latestWaitForCompletionWithTimeoutTaskCount.load());
+    EXPECT_EQ(2u, ultGpgpuCsr->latestWaitForCompletionWithTimeoutTaskCount.load());
     EXPECT_EQ(2u, ultBcsCsr->latestWaitForCompletionWithTimeoutTaskCount.load());
 
     clWaitForEvents(1, &outEvent1);
-    EXPECT_EQ(2u, ultGpgpuCsr->latestWaitForCompletionWithTimeoutTaskCount.load());
+    EXPECT_EQ(1u, ultGpgpuCsr->latestWaitForCompletionWithTimeoutTaskCount.load());
     EXPECT_EQ(1u, ultBcsCsr->latestWaitForCompletionWithTimeoutTaskCount.load());
 
     clReleaseEvent(outEvent1);

shared/source/command_stream/command_stream_receiver.h

@@ -66,7 +66,6 @@ class TimestampPackets;
 
 template <typename T1>
 class TagAllocator;
-class TagNodeBase;
 
 enum class DispatchMode {
     deviceDefault = 0,          // default for given device
@@ -476,7 +475,6 @@ class CommandStreamReceiver {
 
     uint32_t getRequiredScratchSlot0Size() { return requiredScratchSlot0Size; }
     uint32_t getRequiredScratchSlot1Size() { return requiredScratchSlot1Size; }
-    virtual bool submitDependencyUpdate(TagNodeBase *tag) { return true; }
 
   protected:
     void cleanupResources();

shared/source/command_stream/command_stream_receiver_hw.h

@@ -199,7 +199,6 @@ class CommandStreamReceiverHw : public CommandStreamReceiver {
     void programStateBaseAddressHeapless(Device &device, LinearStream &commandStream);
     void programComputeModeHeapless(Device &device, LinearStream &commandStream);
     void handleAllocationsResidencyForflushTaskStateless(const IndirectHeap *dsh, const IndirectHeap *ioh, const IndirectHeap *ssh);
-    bool submitDependencyUpdate(TagNodeBase *tag) override;
 
   protected:
     void programPreemption(LinearStream &csr, DispatchFlags &dispatchFlags);

shared/source/command_stream/command_stream_receiver_hw_base.inl

@@ -2337,25 +2337,4 @@ inline void CommandStreamReceiverHw<GfxFamily>::chainCsrWorkToTask(LinearStream
     this->makeResident(*chainedBatchBuffer);
     EncodeNoop<GfxFamily>::alignToCacheLine(commandStreamCSR);
 }
-template <typename GfxFamily>
-bool CommandStreamReceiverHw<GfxFamily>::submitDependencyUpdate(TagNodeBase *tag) {
-    auto ownership = obtainUniqueOwnership();
-    auto expectedSize = sizeof(typename GfxFamily::PIPE_CONTROL);
-    auto &commandStream = getCS(expectedSize);
-    auto commandStreamStart = commandStream.getUsed();
-    auto cacheFlushTimestampPacketGpuAddress = TimestampPacketHelper::getContextEndGpuAddress(*tag);
-    PipeControlArgs args;
-    args.dcFlushEnable = MemorySynchronizationCommands<GfxFamily>::getDcFlushEnable(true, this->peekRootDeviceEnvironment());
-    MemorySynchronizationCommands<GfxFamily>::addBarrierWithPostSyncOperation(
-        commandStream,
-        PostSyncMode::immediateData,
-        cacheFlushTimestampPacketGpuAddress,
-        0,
-        this->peekRootDeviceEnvironment(),
-        args);
-
-    auto submissionStatus = this->flushSmallTask(commandStream, commandStreamStart);
-    this->latestFlushedTaskCount = taskCount.load();
-    return submissionStatus == SubmissionStatus::success;
-}
 } // namespace NEO

shared/source/command_stream/csr_deps.h

@@ -8,8 +8,6 @@
 #pragma once
 #include "shared/source/utilities/stackvec.h"
 
-#include <set>
-
 namespace NEO {
 
 class TimestampPacketContainer;
@@ -29,7 +27,5 @@ class CsrDependencies {
     void makeResident(CommandStreamReceiver &commandStreamReceiver) const;
     void copyNodesToNewContainer(TimestampPacketContainer &newTimestampPacketContainer);
     void copyRootDeviceSyncNodesToNewContainer(TimestampPacketContainer &newTimestampPacketContainer);
-
-    std::set<CommandStreamReceiver *> csrWithMultiEngineDependencies;
 };
 } // namespace NEO

shared/source/helpers/timestamp_packet_container.h

@@ -39,7 +39,6 @@ struct TimestampPacketDependencies : public NonCopyableClass {
     TimestampPacketContainer barrierNodes;
     TimestampPacketContainer auxToNonAuxNodes;
     TimestampPacketContainer nonAuxToAuxNodes;
-    TimestampPacketContainer multiCsrDependencies;
 
     void moveNodesToNewContainer(TimestampPacketContainer &timestampPacketContainer);
 };